Agent infrastructure
Agent infrastructure is the runtime layer that agents need beyond a model and a framework: isolated sandboxes for executing generated code, cloud browsers for acting on websites, memory services for recalling past interactions, and durable execution engines that keep long-running tasks alive through failures.
Code sandboxes such as E2B, Daytona, Modal Sandboxes, and Vercel Sandbox start isolated machines on demand, while microsandbox and the Kubernetes Agent Sandbox project run on infrastructure you control. Browser providers such as Browserbase, Steel, and Browser Use host headless browsers for agents. Memory layers such as Mem0, Zep, and Cognee extract facts from conversations and retrieve them in later sessions.
Durable execution engines, including Temporal, Inngest, Restate, and DBOS, persist each step of a workflow so an agent loop can pause, retry, and resume. The large clouds bundle several of these pieces into managed agent runtimes: Amazon Bedrock AgentCore, Microsoft Foundry Agent Service, and Google Cloud's Agent Runtime.
Agent infrastructure compared
| Platform | Open source | Deployment | Best for |
|---|---|---|---|
| Agent Runtime (Gemini Enterprise Agent Platform) | No | Managed cloud | Deploying ADK or LangGraph agents on Google Cloud |
| Agent Sandbox (Kubernetes) | Yes (Apache-2.0) | Self-hosted | Platform teams running agent sandboxes on Kubernetes |
| Amazon Bedrock AgentCore | No | Managed cloud | Running framework-agnostic agents on AWS |
| Anchor Browser | No | Managed cloud | Enterprise automation of web-based operations |
| Blaxel | No | Managed cloud | Agents that pause and resume sandboxes often |
| Browser Use | Yes (MIT) | Library, CLI, Managed cloud | LLM-driven web agents in Python |
| Browserbase | No | Managed cloud | Cloud browsers for agents that need the live web |
| Browserless | No | Managed cloud, Self-hosted | Puppeteer and Playwright workloads at scale |
| Cloudflare Sandboxes | No | Managed cloud | Agents built on Cloudflare Workers |
| Cognee | Yes (Apache-2.0) | Library, Self-hosted, Managed cloud | Graph-based agent memory you can self-host |
| Daytona | No | Managed cloud | Stateful sandboxes for coding and agent workloads |
| DBOS | Yes (MIT) | Library | Durable agents on an existing Postgres database |
| E2B | Yes (Apache-2.0) | Managed cloud, Self-hosted | Running agent-generated code in isolated machines |
| Hatchet | Yes (MIT) | Managed cloud, Self-hosted | Task queues and durable agent workflows together |
| Hyperbrowser | No | Managed cloud | Browser sessions and sandboxes from one provider |
| Inngest | No | Managed cloud, Self-hosted | Event-driven agent steps in serverless apps |
| Kernel | No | Managed cloud | Agents that need logged-in, reusable browser sessions |
| LangMem | Yes (MIT) | Library | LangGraph agents that need long-term memory |
| Lightpanda | Yes (AGPL-3.0) | Self-hosted, Local | Lightweight headless browsing at high concurrency |
| Mem0 | Yes (Apache-2.0) | Library, Self-hosted, Managed cloud | Personalized agents that remember users across sessions |
| microsandbox | Yes (Apache-2.0) | Local, Self-hosted | Local, hardware-isolated sandboxes for agents |
| Microsoft Foundry Agent Service | No | Managed cloud | Enterprise agents on Azure with Entra identity |
| Modal Sandboxes | No | Managed cloud | High-volume agent and reinforcement learning workloads |
| Northflank Sandboxes | No | Managed cloud | Teams already deploying workloads on Northflank |
| Restate | No | Self-hosted | Durable agent logic with low-latency state |
| Steel | Yes (Apache-2.0) | Managed cloud, Self-hosted | Self-hostable browser sessions for agents |
| Supermemory | No | Managed cloud, API | Adding user memory to agents through an API |
| Temporal | Yes (MIT) | Self-hosted, Managed cloud | Long-running agents that must survive failures |
| Trigger.dev | Yes (Apache-2.0) | Managed cloud, Self-hosted | TypeScript teams running long background agent tasks |
| Vercel Sandbox | No | Managed cloud | Apps on Vercel that execute generated code |
| Workflow SDK | Yes (Apache-2.0) | Library | Durable agent steps in TypeScript web frameworks |
| Zep | No | Managed cloud | Agents that need memory of facts that change over time |
Profiles

Agent Runtime (Gemini Enterprise Agent Platform)
Google Cloud · docs.cloud.google.com
Agent Runtime is the managed agent hosting service in Google Cloud's Gemini Enterprise Agent Platform, deploying ADK, LangGraph, LangChain, AG2, and custom agents with sessions, Memory Bank, sandboxed code execution, and evaluation services.

Agent Sandbox (Kubernetes)
Kubernetes SIG Apps · agent-sandbox.sigs.k8s.io
Agent Sandbox is a Kubernetes SIG Apps project that adds a Sandbox custom resource and controller for isolated, stateful, singleton workloads such as agent runtimes, delegating isolation to runtimes like gVisor or Kata Containers.

Amazon Bedrock AgentCore
Amazon Web Services · aws.amazon.com
Amazon Bedrock AgentCore is a set of AWS services for running agents built with any framework and model, including a serverless runtime with session isolation, a managed agent runtime, memory, gateway, identity, browser, code interpreter, and observability.

Anchor Browser
anchorbrowser.io
Anchor provides secure browser infrastructure for computer-use AI agents, letting teams deploy fleets of cloud browser agents that automate web tasks otherwise done by people in enterprise operations.
Blaxel
blaxel.ai
Blaxel is infrastructure for autonomous agents built on microVM sandboxes that boot and resume in milliseconds, with shared memory and programmable networking for agents that run code and tools between long idle periods.

Browser Use
browser-use.com
Browser Use is an open-source Python library for browser agents that operate websites with any LLM, plus a cloud service that provides hosted web agents and stealth browsers that other agent frameworks can connect to.
Browserbase
browserbase.com
Browserbase runs headless browsers in the cloud for AI agents and automations, and it maintains Stagehand, an MIT-licensed framework for browser automations that combine code with natural-language instructions.
Browserless
browserless.io
Browserless runs cloud browsers for AI agents and automation that connect over MCP, Puppeteer, or Playwright, with stealth and authenticated profiles; its self-hostable code is source-available under SSPL-1.0 or a commercial license.

Cloudflare Sandboxes
Cloudflare · developers.cloudflare.com
Cloudflare Sandboxes execute untrusted or generated code in Linux virtual machines or isolated Workers, managed from Workers through Cloudflare's open-source Sandbox SDK, for agents that already run on Cloudflare's developer platform.
Cognee
cognee.ai
Cognee is an open-source memory platform for LLM agents that builds persistent memory across sessions with graph, vector, and relational retrieval, running self-hosted, in Docker, on-premises, or on Cognee Cloud.
Daytona
daytona.io
Daytona runs secure, elastic sandboxes for AI-generated code with stateful operations and SDKs for agents; in June 2026 Daytona moved core development to a private codebase and stopped updating its formerly open-source GitHub repository.

DBOS
dbos.dev
DBOS is an open-source durable workflow library built on Postgres: developers annotate workflows and steps in their code to make them durable, with no separate orchestrator to operate, for agent loops, queues, and scheduled jobs.

E2B
e2b.dev
E2B provides isolated sandboxes in which AI agents run code, browse, and use tools, with pause and fork support; its SDKs and infrastructure are open source, and sandboxes run in E2B's cloud or a customer's AWS, GCP, or Azure account.
Hatchet
hatchet.run
Hatchet is an open-source orchestration engine for background tasks, AI agents, and durable workflows, with SDKs for Python, TypeScript, Go, and Ruby, available as Hatchet Cloud or self-hosted.
Hyperbrowser
hyperbrowser.ai
Hyperbrowser provides cloud browsers, isolated sandboxes, and web data APIs for AI agents, starting Chrome sessions and microVMs through one API that Playwright, Puppeteer, or agent frameworks connect to; its HyperAgent library is AGPL-licensed.
Inngest
inngest.com
Inngest is a durable execution platform for background jobs, workflows, and AI agents, where step functions written with its SDKs get retries, flow control, and step-level observability; its server is source-available under SSPL with a delayed Apache-2.0 license.
Kernel
kernel.sh
Kernel is browser infrastructure for AI agents and web automations, with autoscaling cloud browsers, reusable sessions, and managed authentication, and it publishes its browser container images as open source under Apache-2.0.

LangMem
LangChain · langchain-ai.github.io
LangMem is LangChain's open-source Python library for agent long-term memory, with tools agents call to record and search memories, a background manager that extracts and consolidates knowledge, prompt refinement, and native LangGraph store integration.

Lightpanda
lightpanda.io
Lightpanda is an open-source headless browser written from scratch in Zig for machines rather than people, designed for web automation and AI agents with a smaller memory footprint than headless Chrome.
Mem0
mem0.ai
Mem0 is a memory layer for AI agents and applications that extracts, stores, and retrieves facts from past interactions so agents can personalize responses, available as an Apache-2.0 open-source library and a managed platform.

microsandbox
microsandbox.dev
microsandbox is an open-source microVM sandbox for AI agents and untrusted code that runs locally on macOS, Linux, and Windows or on self-hosted servers, with access to a managed cloud version available on request.

Microsoft Foundry Agent Service
Microsoft · learn.microsoft.com
Microsoft Foundry Agent Service hosts prompt agents, voice agents, and containerized hosted agents built with Agent Framework, LangGraph, or other SDKs, with toolboxes exposed over MCP, Microsoft Entra agent identities, tracing, and A2A endpoints.
Modal Sandboxes
Modal Labs · modal.com
Modal Sandboxes let developers create isolated environments programmatically on Modal's serverless compute platform to run agent code and reinforcement learning rollouts on any container image, with CPU and memory that burst on demand.
Northflank Sandboxes
Northflank · northflank.com
Northflank Sandboxes are isolated microVM environments for running untrusted and AI-generated code at scale, offered as part of Northflank's cloud platform for deploying and operating application workloads.
Restate
restate.dev
Restate is a durable runtime for agents and backends that packs durable execution, keyed state, queues, RPC, timers, and flow control into a single binary; the server is source-available under the Business Source License 1.1.
Steel
steel.dev
Steel is an open-source browser API for AI agents that manages sessions, pages, and Chrome processes, with control through Puppeteer, Playwright, or Selenium over CDP, available as a self-hosted server or a managed cloud service.
Supermemory
supermemory.ai
Supermemory is a memory and context engine for AI agents that extracts facts from conversations, maintains user profiles, handles contradictions and expiring information, and adds retrieval and connectors, offered through an API, plugins, and MCP.
Temporal
Temporal Technologies · temporal.io
Temporal is an open-source durable execution platform whose server runs workflows that survive process crashes and outages by persisting each step and retrying failures, available self-hosted or as Temporal Cloud and used for long-running agent loops.
Trigger.dev
trigger.dev
Trigger.dev is an open-source platform for durable AI agents and workflows in TypeScript that runs long-running tasks with retries, queues, and LLM observability, available as a managed cloud or self-hosted.

Vercel Sandbox
Vercel · vercel.com
Vercel Sandbox runs AI-generated, customer, or other untrusted code in isolated, ephemeral virtual machines on Vercel's infrastructure, controlled through an Apache-2.0 SDK from Vercel functions or other environments.
Workflow SDK
Vercel · workflow-sdk.dev
Workflow SDK is Vercel's open-source TypeScript library that makes async functions durable through a use workflow directive, so apps and agents can suspend, resume, and keep state, with deployment on Vercel, other clouds, or Docker.

Zep
getzep.com
Zep is a context layer for agents that builds a temporal knowledge graph from conversations and business data using its open-source Graphiti library; Zep's self-hosted Community Edition is deprecated, so Zep itself is offered as Zep Cloud.
How to choose
- 01Isolation model
- MicroVMs, hardened container runtimes such as gVisor or Kata Containers, and plain containers give different security boundaries for untrusted, model-generated code.
- 02Session lifecycle
- Start time, maximum session length, and pause, resume, snapshot, or fork support determine whether an agent can keep state between steps.
- 03Deployment and data location
- Managed services are quickest to adopt, while self-hosted or bring-your-own-cloud options keep code, browser sessions, and memories inside your own network.
- 04Licensing
- Some entries are open source, while others are source-available under licenses such as SSPL or the Business Source License that restrict offering the software as a competing service.
- 05SDKs and integrations
- Official SDK languages and ready-made integrations with agent frameworks, Playwright or Puppeteer, and MCP reduce glue code.
- 06Observability and controls
- Session recordings, step-level traces, audit logs, and network egress controls help teams debug agents and enforce security policy.
Frequently asked questions
What is an AI agent sandbox?
- An AI agent sandbox is an isolated computing environment, usually a microVM or hardened container, where an agent can run code, install packages, and edit files without access to the host system or other tenants. Sandboxes are created on demand through an API, can often be paused or snapshotted, and are destroyed when the task ends. E2B, Daytona, and Modal Sandboxes are examples.
What is agent memory in AI?
- Agent memory is the set of mechanisms that let an AI agent keep information beyond a single context window, such as user preferences, facts learned in earlier sessions, and summaries of past tasks. Memory layers extract these items from conversations, store them in vector, graph, or relational databases, and retrieve the relevant ones into later prompts. Mem0, Zep, Cognee, and LangMem are examples.
What is durable execution?
- Durable execution is a programming model in which an engine records the result of each workflow step, so a program that crashes, times out, or waits days for human approval resumes from the last completed step instead of starting over. Agents use it for long-running loops with many model and tool calls. Temporal, Inngest, Restate, DBOS, and Trigger.dev provide durable execution.
E2B vs Daytona: what is the difference?
- Both services start isolated sandboxes for AI-generated code through SDKs. E2B publishes its SDKs and sandbox infrastructure under Apache-2.0 and can run sandboxes in E2B's cloud or in a customer's AWS, GCP, or Azure account. Daytona offers managed, stateful sandboxes; in June 2026 it moved core development to a private codebase and stopped updating its formerly open-source repository.
Related guides
AI agents hubOther categories
- Model providers
- Inference providers
- GPU clouds
- Serving engines
- Vector databases
- Embedding and reranking models
- Document parsing and data ingestion
- Fine-tuning and post-training
- Agent frameworks
- Web search APIs
- MCP platforms
- Voice AI
- AI gateways
- LLM observability and evaluation
- Guardrails and AI security
- AI coding tools
Product details change quickly; check each vendor’s site for current features, licensing and pricing.